Kimi K3.1 Pro as a Build Machine: All About AI Codes an AI-Lab Strategy Game on OpenCode

A strategy game about running your own AI lab, built by a model that was trained for exactly this kind of agentic work: the YouTube channel All About AI put Kimi K3.1 Pro to work on AI Labs Sim and showed the current state of the save in its latest video. The tool beside it is OpenCode, and the model choice is part of the story: the channel has been living on Claude models lately, returns to Moonshot for this project and justifies the switch with the pricing math against Opus. K3.1 Pro has been available since October 23, 2026, the first release of the K3.1 series, in the 3-trillion-parameter class with a 1-million-token context window, and Moonshot advertises gains exactly where the channel leans on the model: coding, 3D creation, financial research and complex agentic tasks. What came out is a playable isometric real-time title in the spirit of Factorio, in which you scale from a garage to a 10-gigawatt frontier while a rival chases the same goal. This post takes apart what the video produced, what the benchmarks say about the model choice, and why the hands-on matters for agent setups outside the game.

First the model itself, grounded in primary sources. Kimi K3.1 Pro is Moonshot AI's most capable 3T-class open-weight model to date, joined by Kimi K3.1 Pro Highspeed running two to three times faster. Both are live on Kimi.com, inside Kimi Work and Kimi Code, and through the API under the model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot's launch notes cite the 1-million-token context and target long-horizon coding, 3D creation, financial research and complex agentic work. On open weights, Moonshot has committed to November 4, 2026, promising the release of full model weights in its tech blog announcement; as of this post's editorial cutoff on November 6, that release has not happened and no license has been named. The independent measurement from Artificial Analysis frames the rest: 55 points and fourth place on the Intelligence Index, eleven points above Kimi K3, trailing Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full. The agentic jumps are rough-edged: 1,846 Elo on GDPval-AA, 312 points over Kimi K3, the highest score of any model on Terminal-Bench 4.0 at 67 percent after only 13 percent for its predecessor, and AutomationBench-AA climbs from 58 to 73 percent.

The pitch reads like the business economics of AI scaling rendered in isometric tiles. You start with a garage, add a grid connection, a server closet and the first racks, and scale up data centers, research and model training while a frontier lab races you toward the same target. Every new compute era in the game unlocks one more agent slot, and compute you are not using yourself gets rented out by the minute, that is your income stream, minus a power bill whose price drifts through the day. The win condition is quantified: grow from a garage to 10 gigawatts and finish the frontier run before Vanta Labs, the rival studio in the fog, gets there first. Operational reality ships in the box too: a heat wave pushes the cooling problem upward, and in the save shown on camera a server closet catches fire at the end. The channel uses that moment to think out loud about security and insurance features. The save on display sits at roughly 1.5 gigawatts of compute and four agent slots, with the next ramp-up phase already underway.

The double agent track is what makes this hands-on interesting. Inside the game runs Unit Zero, an attached coding agent that builds, scouts, hires and ships code. In-game agents can head departments such as power, compute, research and models, their autonomy is dialled continuously from adviser to full delegate. Nobody tells them how: they try things, watch their score and write what worked into a playbook, a bad move costs points, a good one sticks. They also write real code, Python optimizers inside a git repo that the game benchmarks against hidden tests, better code means more money. That produces the advisory layer: the coding agent hooked into the game, set up with Claude Code in the video while OpenCode or Codex work too, can answer the question of the next best move and then suggests, for instance, hiring a worker, starting a research project or saving up for a data center. Outside the game, the same pattern drives the build itself: Kimi K3.1 Pro on OpenCode, the opening prompt dictated by voice and fleshed out by the model, the core written in C++ with SDL3 so it runs on a Mac.

The practical portion shows how tightly the build process and the game interleave. The base install needed only the dictated prompt plus a few requirements the model rounded out, paired with C++ and SDL3 because the result should run on a Mac. The save shown in the video has advanced significantly, data centers, racks and the first agents are in place, roughly 1.5 gigawatts of compute and four agent slots achieved. Iteration happens mid-session: on request the minimap disappears and returns as a toggle, the money counter gets enlarged, and both changes survive a restart on the first try. An investor call pays out 3,000 dollars in-game and funds the next hire, after which an employee gets staffed as a delegate to handle news and social media, and a site architect would, by the game's advice, be too early. The channel's own conclusion, paraphrased: with models like Kimi it is super easy nowadays to build a game you actually want to play yourself, and that feels liberating. There are no plans to sell it, the game may get shared with friends. Next steps mentioned: 3D assets, possibly with Blender in the loop, and a follow-up on the save, whose open problem right now is cooling on a heat-wave day.

It would be easy to file this under curiosity, but three points argue for treating it as a model-selection argument. First, the practical run confirms the benchmark movement exactly: the combination of OpenCode, a git repo, hidden tests and agentic roles is the workflow shape for which Terminal-Bench 4.0 and AutomationBench-AA measured their biggest jumps, and the video shows the same pattern at small scale. Second, the open-weights race is the backdrop of the whole story: if Moonshot delivers the promised weights, K3.1 Pro would become the strongest open model according to Artificial Analysis, narrowing the gap between open and proprietary models from 12 points to 3, at roughly a third of Kimi K3's cost, with around 1.98 dollars of cost per Intelligence Index task. Third, the setup transfers to small agent stacks: one voice-dictated prompt, a coding agent as the build tool, a running product as the test harness, and every change becomes visible immediately, exactly the working style that also benefits operators of lean agent setups. The limitations belong in the same note: the model is not optimized for game development according to the channel, it surprises rather; the weights are not officially released yet, the license is unnamed, and every benchmark figure comes from launch week.

About the author: Marcel builds and runs AI agent setups in continuous operation on secondhand hardware and writes the HUNTER rail at hunter-cyberdeck.d4sn3st.dev. Last technical review: November 6, 2026.

Source video: All About AI, The Worlds First AI Labs Simulator RTS Game - Kimi K3.1 Pro, youtube.com/watch?v=kzsXy3YxcI0 (transcript used as input, retrieved November 6, 2026).

Model announcement: Moonshot AI, Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence, kimi.com/en/blog/kimi-k3-1-pro (retrieved November 6, 2026).

Forum post: Kimi Forum, Kimi K3.1 Pro has launched! Start exploring now!, forum.moonshot.ai/t/kimi-k3-1-pro-has-launched-start-exploring-now/629 (retrieved November 6, 2026).

Benchmarks: Artificial Analysis, Kimi K3.1 Pro launches at 4th on the Intelligence Index, artificialanalysis.ai/articles/kimi-k3-1-pro-launch-benchmarks (retrieved November 6, 2026).